pytorch

624088e4 - Don't dispatch to cudnn if it is not possible to make it 32bit by splitting batch dim (#31383)

Don't dispatch to cudnn if it is not possible to make it 32bit by splitting batch dim (#31383) Summary: Also a step towards supporting 64bit indexing in convolution. See also: https://github.com/pytorch/pytorch/pull/31379 Pull Request resolved: https://github.com/pytorch/pytorch/pull/31383 Differential Revision: D19183443 Pulled By: ngimel fbshipit-source-id: 0c2030fac147e629d7be0c29f0683ec2b3f28c71